DeepSeek releases Agent model at the end of the year, with Liang Wenfeng making a heavy bet

Bloomberg reported that DeepSeek is developing a new large-scale model with stronger autonomy, aiming to directly compete with American giants such as OpenAI in the next wave of AI agent competition.

The previous large models relied on users repeatedly giving instructions to work, but this new model is different. Users only need to give a little hint, and it can complete multi-step tasks on its own, and continuously learn and optimize during the process of doing things.

According to insiders, this indicates that DeepSeek is going to turn its product into a “true intelligent agent”, which is the next stage of artificial intelligence development. Founder Liang Wenfeng plans to launch this new software in the fourth quarter of this year, and this timing is crucial. The entire industry is now waiting for R1’s successor.

At the beginning of the year, R1 went viral, costing millions of dollars, but its results were able to match or even surpass OpenAI, directly shaking the global technology community. DeepSeek was also immediately regarded as the most likely AI new company in China to disrupt the industry. But in the following months, DeepSeek only made some minor updates, while its Chinese and American counterparts kept releasing new models. The outside world guessed why R2 was delayed. There are two theories: one is that Liang Wenfeng wants to do his best and doesn’t want to pass on tasks casually; Secondly, there were training and technical difficulties encountered during the research and development process.

In fact, DeepSeek’s new plan is not an isolated case, as global tech giants are turning towards AI agents. In recent months, OpenAI, Anthropic, and Microsoft have all launched their own Agent software, aiming to change the way people work and live. There is also a Chinese startup called Manus AI, which has become popular worldwide with the concept of “universal agents”.

Compared to early chatbots that could only reply with short texts, people hope that this generation of products can handle more complex practical tasks. From planning trips to writing and debugging code, truly achieving “do it yourself”, but most of the agents on the market still rely on human monitoring, far from completely doing their own work. DeepSeek wants to do as much as possible without human “monitoring”, so that AI can truly act for people reliably.

Compared to local giants such as Alibaba and Tencent, DeepSeek’s research and development pace is much slower. Alibaba’s Qwen series models have been attracting users, and Tencent frequently releases new versions. However, DeepSeek has always been low-key and unhurried, which is actually Liang Wenfeng’s strategic choice: before the new form of Agent matures, it will not engage in frequent updates and focus resources on products that can truly change the industry landscape. It is still uncertain whether DeepSeek can once again make the world discuss it like R1 did, but what is certain is that the company is betting its future on the “AI Agent” track, which is also the next key battlefield for the entire artificial intelligence industry.